Speech event separated method, speech event separated system, and speech event separation program

摘要

P To provide a method, a system and a PROGRAM for utterance event separation that can separate utterance contents by speakers with high precision while excluding overlap of utterance during a conference. PSOLUTION: Sound source directions by times are estimated from multi-channel voice data recorded in a continuous section of the conference during the conference. Presence ranges of speakers as sound sources are estimated to identify who speaks by the times. Consequently, a position vector of a target speaker is estimated and a noise space correlation matrix of other speakers is calculated to generate a filter based upon the position vector of the target speaker and the noise space matrix of the other speakers with respect to the target speaker and the filter is used to separate and output an utterance of only the target speaker. PCOPYRIGHT: (C)2007 and JPO& INPIT